Twitxy Saved

Shubh Sinha

Verified account This account is verified. Learn more

@SHUBH17_

NYC Joined November 2011
586 Following 1.1K Followers 486 Posts

Latest posts 5

Pinned post

Shubh Sinha Verified account @SHUBH17_
馃帀I鈥檓 excited to announce our $18M Series A, with participation from @VenrexVC, @thegp , @VirtueVC , @caffeinatedcap , @arrayvc, @GreatPointVC , @LiveRamp, @haystackvc, @CapitalAlso, @LifeXvc , Circle & Co, and WS Investments.

The first generation of AI was built on public, human-curated data. The next will be built on real-world data. The datasets carrying real human behavior are what now separate one AI system from another. AI companies want them. The companies that hold them want to put them to work. The barrier has always been risk.

We solved that risk in healthcare, one of the most regulated data environments there is. Today we bring it to the rest of the economy.
1:57
50 replies 30 reposts 346 likes 110K views View on X
Shubh Sinha Verified account @SHUBH17_
If data becomes one of the world鈥檚 most valuable traded assets then it鈥檚 fair to say it will need better market infrastructure. Every working market needs the below:

Independent evaluators. Liquidity for any asset class is unlocked when there is trust established around the contents without having to check firsthand. Commodities, energy, equities, and debt all cross that threshold through grading and verification by someone who isn't the seller.

Transparency. The biggest markets moved beyond one-off negotiations once the terms became open and public: what was traded, at what price, under what conditions. Data deals today are almost all private, so there's no market-clearing price for anyone to point to.

Liquidity. That takes sufficient scale of buyer and seller transactions, and it only shows up once the first two exist. Buyers who can trust a dataset's grade and see what comparable datasets went for can transact a lot quicker.

Data has ridiculous demand but has not yet met the three conditions.
2 replies 0 reposts 15 likes 1.1K views View on X
Shubh Sinha Verified account @SHUBH17_
Data vendors still have to navigate the fact that de-id is almost always self-attested.

It鈥檚 somewhat of a conflict of interest as the party checking the data will benefit when the deal closes. So there's a baked-in incentive to say de-id will be done to best-in-class benchmarks, whether or not anyone outside the deal can see how.

For a lot of smaller deals that might be workable. For the most complex enterprise-level deals, independent review and transparency are becoming necessary as the residual risk and external scrutiny make self-attestation insufficient. We saw this with Spirit Airlines, where the sale required an independent de-id agent to sign off with the court before any data moved, and an independent privacy ombudsman was appointed.

A great parallel is financial advisory. The stakes also become high enough that self-attestation loses some credibility. Independent advisors have been taking market share from bigger banks as they don鈥檛 have a stake in the outcome and aren't incentivized through financing interest.
1 replies 0 reposts 10 likes 879 views View on X
Shubh Sinha Verified account @SHUBH17_
Naive de-identification is a very expensive shortcut in agent training data.

This typically includes editing data with basic regex or random token masking to replace anything that looks sensitive (names, personal details etc) with a generic placeholder. Because it doesn't read the surrounding context, the same person that appears again later in the trajectory might now have a different placeholder.

Outputs look clean because the obvious identifiers are gone. But the dataset is now internally inconsistent, i.e. the causal graph is broken. Any agent training on this cannot learn which action caused which observation and RL loses its effectiveness.

Incredibly important to retain the signal / ensure the entity graph doesn鈥檛 break. We're early in how data gets prepared, and I think how well de-id preserves that graph will decide how much enterprise trajectory data is actually usable.
2 replies 0 reposts 11 likes 702 views View on X
Shubh Sinha Verified account @SHUBH17_
Companies need to think about how to navigate the market for data.

One of the first decisions is to figure out which data stays internal and which is shared externally.

This is a data allocation decision, likened to how CFOs decide what to return to shareholders vs. let compound. If we assume companies can de-id data while retaining high-quality signal then there will be a choice to make.

The stuff that leaves is the data that has more value to an external buyer (lab or model trainer) than it has residual strategic value to said company. That tends to be historical rather than ongoing production data, and it likely comes from non-core parts of the business. An industrial company鈥檚 accounting workflows, back-office processes, or older operational logs are not necessarily the trade secrets.

Until recently this wasn't a choice at all. Almost everything sat locked up behind obligations but it potentially becomes a key CEO-level decision over the coming years.